The clock, used for trace time stamps, can vary between CPUs.
When synchronizing host and guest trace time stamps, these
variations could introduce inaccuracy.
The algorithm for host - guest trace time stamps synchronization
is modified to calculate the offset per each host CPU, depending
on the CPU affinity of host tasks, running the guest's VCPUs.
The calculated array per CPU is stored in the trace.dat file, using
TRACECMD_OPTION_TIME_SHIFT option.
The "trace-cmd dump --options" command and tracecmd_init_data() API
are adjusted to read the new format of this option.
In order this per CPU synchronization data to be used when host and
guest files are merged, information about guest VCPUs migration
between host CPUs is needed. This information is extracted from
host trace data, thus the host trace must be recored with "-e sched"
option.
A new API tracecmd_open_merge() is introduced, which triggers this
per CPU calculations when opening tracing files. This API must be
used by KernelShark when opening host and guest files. There is a
separate patch for this, not part of the series.
Tzvetomir Stoyanov (VMware) (5):
trace-cmd: Time stamp offset per host CPU
trace-cmd: Fix reading of the traceid option from trace.dat file
trace-cmd: Add new API to merge two trace files
trace-cmd: Add a define to enable per CPU timestamps synchronization
trace-cmd: Use per CPU synchronization data when calculating
timestamps offsets
include/trace-cmd/trace-cmd.h | 7 +-
lib/trace-cmd/include/trace-tsync-local.h | 16 +-
lib/trace-cmd/trace-input.c | 516 +++++++++++++++++++---
lib/trace-cmd/trace-timesync.c | 269 +++++++++--
tracecmd/include/trace-local.h | 4 +
tracecmd/trace-dump.c | 60 ++-
tracecmd/trace-record.c | 18 +
tracecmd/trace-tsync.c | 144 +++---
8 files changed, 832 insertions(+), 202 deletions(-)
--
2.25.1
|